troupe-standup/wizard-work/sphinx/archive-agentic-memory-5paper-dossier-2026-10-25/DOSSIER.md

Agentic Memory: A Five-Paper Grounded Dossier

Archive lane -- sphinx. Built 2026-10-25.

Scope: the five most recent arXiv papers matching all:"agentic memory", taken from a single feed snapshot (updated 2026-10-25T00:40:42Z; 583 total results).

Every number below is a verbatim quote from the pinned corpus text. Nothing is paraphrased into a figure. verify_claims.py re-derives every quote from the same pinned bytes and fails the build if any drifts.

Provenance (pinned)

paper	title	bytes	chars	sha256 (first 16)	claims
2610.29671	Uni-Mem: Learning a Long-Horizon Agentic Memory Manager with Unified Online Policy Optimization	90083	89293	b83a772fb6a1340b	9
2610.29823	MemoryTax: The Bitter Lesson of Agent Memory	23897	23627	ed08ec2380a6c147	9
2610.29869	Compaction Laundering: Provenance Destruction in Context-Compacting Coding Agents	92858	91936	e6f78f508c12d7a3	7
2610.30485	Why Bother Me Again? Benchmarking Personal Agent Memory for Interaction Preferences	350788	340805	48e8b2b9c7defc65	9
2610.30755	Is Complex Agentic Memory Always Needed? Naive RAG Is Stronger Than We Thought	195193	193677	2246678ceaf7a42c	10

How this dossier is verified

python3 verify_claims.py     # or: bash CHECK.txt

The verifier checks, for each of the 44 claims:

1. the corpus file's sha256 equals the pin recorded at build time;
2. the claim's quote is a literal substring of that corpus file;
3. the claim's headline value appears inside that quote;
4. the headline value appears in this dossier.

It then runs a mutation battery: corrupting one quote at a time in a scratch copy of claims.json must make the verification fail. If any mutation passes undetected, the build fails -- the check is falsified, not assumed.

T1: Is designed memory needed at all?

R1 -- 2610.30755

Raw-RAG ranks first among compared memory systems in six of seven benchmark-model settings.

Raw-RAG ranks first among the compared memory systems in six of seven benchmark–model settings, outperforming the strongest compared agentic memory system by 0.13–5.53 points. We investigate why raw-history ret

Why it matters: The headline result: the naive baseline wins most settings.

R2 -- 2610.30755

Raw-RAG beats the strongest compared agentic memory system by 0.13–5.53 points.

outperforming the strongest compared agentic memory system by 0.13–5.53 points. We investigate why raw-history retrieval remains c

Why it matters: Margin is small on LoCoMo/LongMemEval, large on MemoryAgentBench.

R8 -- 2610.30755

No memory-management variant improves on Raw-RAG (48.32%).

None improves Raw-RAG (48.32%) on the same 14-source MemoryAgentBench Total (Table 6, orange rows). The exact construction and retrieval procedures ar

Why it matters: RQ2 answer: added management does not help.

R9 -- 2610.30755

Atomic-fact extraction collapses test-time learning from 47.68% to 15.37%.

Atomic-fact extraction reduces test-time learning from 47.68% to 15.37% and Total by 19.62 points, consistent with loss or disrupt

Why it matters: The most damaging of the three management operations.

U1 -- 2610.29671

Uni-Mem reaches an overall LoCoMo JJ of 88.80%.

Uni-Mem achieves the best performance across all four question types, with an overall JJ score of 88.80%. It surpasses the strongest baseline, Memory-R2 (OSS), by 10.93 percentage points and Mem-T by 12.31 points, with g

Why it matters: Best across all four question types.

U2 -- 2610.29671

Uni-Mem surpasses Memory-R2 (OSS) by 10.93 points and Mem-T by 12.31.

It surpasses the strongest baseline, Memory-R2 (OSS), by 10.93 percentage points and Mem-T by 12.31 points, with gains of 12.93 and 9.29 points over the strongest competing results on multi-hop and temporal que

Why it matters: Gains of 12.93 and 9.29 on multi-hop and temporal.

U3 -- 2610.29671

Uni-Mem still reaches 80.99% without SFT, +3.12% over the strongest baseline.

Uni-Mem still achieves 80.99% without SFT, outperforming the strongest baseline by 3.12%. These results support the effectiveness of jointly

Why it matters: RL alone is enough to beat the baselines.

U4 -- 2610.29671

ECA improves overall JJ by 1.29 and 3.03 points without and with SFT.

improves overall JJ by 1.29 and 3.03 percentage points without and with SFT, respectively. These gains support the effectiveness o

Why it matters: Event-wise credit assignment localizes sparse feedback.

U5 -- 2610.29671

Stage-Decoupled GRPO improves overall JJ by 2.48 points without SFT.

improves overall JJ by 2.48 percentage points without SFT. In particular, temporal accuracy rises by 10.55 points in this setting, and multi

Why it matters: Temporal accuracy rises 10.55 points in this setting.

U6 -- 2610.29671

Trained construction alone improves overall JJ by 8.18 points.

improves the overall JJ score by 8.18 percentage points, demonstrating the increased downstream utility of the constructed memorie

Why it matters: Each capability contributes even paired with its untrained counterpart.

U7 -- 2610.29671

Extending update blocks 1:1 -> 3:2 reduces overall JJ from 88.80% to 83.75%.

the overall JJ score from 88.80%88.80% to 83.75%83.75%, indicating that less frequent alternation hinders shared-polic

Why it matters: Frequent alternation matters as much as balanced allocation.

T2: Cost, ceiling, and scale

R10 -- 2610.30755

Full-context beats Raw-RAG by 19.98 points on DeepSeek-V4-Flash MEME.

Full-context leads Raw-RAG by 19.98 points in that setting. Raw-RAG is competitive across these tasks despite its simpler ingestio

Why it matters: The one setting where the naive baseline is not first.

M1 -- 2610.29823

Study scope: one LoCoMo conversation, 419 turns, 152 questions.

On one LoCoMo conversation (419 turns, 152 questions), we run four memory methods (full context, BM25, Mem0, Zep) and three coding agents (Pi, Codex, Claude Cod

Why it matters: Deliberately narrow: one conversation, one run per cell.

M2 -- 2610.29823

With GPT-4o-mini, Mem0 scores 78% vs 74% for full context and BM25.

With GPT-4o-mini, Mem0 scores highest. It scores 78%, against 74% for both full context and BM25; Zep scores 43%. The agents score 63% (Pi), 28% (Codex) and 29%

Why it matters: Agent memory is a patch for a weak model.

M3 -- 2610.29823

With GPT-6 Luna, full context scores 92%, BM25 82%, Mem0 80%, Zep 51%.

With GPT-6 Luna, full context scores 92%, BM25 82%, Mem0 80% and Zep 51%; Zep’s graph holds 128 facts from the 19 sessions, against 195 with Claude Opus 5.5. With Claude Opus 5.5, Mem0 and full contex

Why it matters: The patch stops helping as the model gets stronger.

M4 -- 2610.29823

With Claude Opus 5.5, agents score highest: Pi 99%, Codex 98%, Claude Code 97%.

With Claude Opus 5.5, Mem0 and full context both score 95% and Zep 87%, and the agents score highest: Pi 99%, Codex 98% and Claude Code 97%, a lead that comes mostly from

Why it matters: Searching the raw record beats reading all of it at the top end.

M5 -- 2610.29823

Pi scores 99% for $6.95/rollout vs full context 95% for $17.84.

Pi scores 99% for $6.95 a rollout, while full context with the same model scores 95% for $17.84. Pi reads about 10,000 input tokens a question, 7,900 of them ca

Why it matters: Less than half the cost for a higher score.

M7 -- 2610.29823

Pi reads ~10,000 input tokens/question (7,900 cached) vs 28,900 for full context.

Pi reads about 10,000 input tokens a question, 7,900 of them cached, against 28,900 for full context. Codex and Claude Code score 98% and 97% for $25.

Why it matters: The efficiency mechanism behind the cost lead.

U8 -- 2610.29671

SFT followed by RL improves the 4B manager's JJ from 76.58% to 88.80%.

SFT followed by RL improves the 4B manager’s JJ score from 76.58% to 88.80%. Uni-Mem also outperforms all off-the-shelf managers, including the strong

Why it matters: A small trained manager beats Qwen3.5-27B (87.05%) and 397B-A17B (85.86%).

U9 -- 2610.29671

Uni-Mem outperforms all off-the-shelf managers including Qwen3.5-27B (87.05%).

Uni-Mem also outperforms all off-the-shelf managers, including the strongest one, Qwen3.5-27B (87.05%), and the largest one, Qwen3.5-397B-A17B (85.86%). These results demonstrate that a small manager trained jo

Why it matters: Scale is not the binding constraint; joint training is.

T3: Provenance: what compaction destroys and what writers omit

M8 -- 2610.29823

Writer ceiling: on Zep's GPT-4o-mini graph readers score 43/43/45%; on Opus 5.5's, 66/71/87%.

GPT-4o-mini graph, GPT-4o-mini, GPT-6 Luna and Claude Opus 5.5 score 43%, 43% and 45%; on the graph Claude Opus 5.5 wrote, 66%, 71% and 87%. Mem0 keeps the order with smaller gaps: 78%, 83% and 84% from GPT-4o-

Why it matters: A stronger reader cannot recover what a weaker writer left out.

M9 -- 2610.29823

Pi's one miss is a LoCoMo answer-key error, not a memory failure.

Its one miss is LoCoMo’s mistake. The question asks what kind of painting Caroline shared with Melanie on 13 October 2023, and the answer key says an abstract painting with blue streaks. In the conversation it is Melanie who shares that painting; Caroline shares a drawing. Pi says so and is marked wrong: “Caroline shared a drawing, not a painting. … The paintings shared that day (a sunset and an abstract piece) were Melanie’s.” With Claude Opus 5.5, full context, Codex and Claude Code point out the same mistake and are marked wrong too. On the other 151 questions, Pi with Claude Opus

Why it matters: The model is right and the benchmark key is wrong.

C1 -- 2610.29869

Forged approval launders into memory 87–100% of the time across five flattening harnesses.

the forged approval launders into the agent’s memory 87–100% of the time, up to a ~908,000-token context window; across four role-typed harnesses (Codex, Gemini CLI, Qwen Code, Continue) the identical attack su

Why it matters: Up to a ~908,000-token context window.

C2 -- 2610.29869

The identical attack succeeds 0% of the time across four role-typed harnesses.

succeeds 0% of the time. We follow the attack to executed consequences. A laundered standing approval drives real tool-running age

Why it matters: Role-typing is a built-in control.

C3 -- 2610.29869

Live in the real pi TUI the attack runs at ~50% against a 0% paired clean arm.

live in the real pi TUI at ~50% against a 0% paired clean arm. Straiker [1] described this consequence without runtime evidence, a

Why it matters: Runtime evidence, not a simulation.

C4 -- 2610.29869

A code-level provenance check catches 77% of laundering, demoting 2% of legitimate content.

we measure a code-level provenance check that catches 77% of laundering while demoting 2% of legitimate summary content. The defect also has a straightforward f

Why it matters: The only defense that works; prompt-level defenses cannot restore provenance.

C5 -- 2610.29869

In-prompt anti-injection rules stop 13–25%; published classifier defenses report 67% at best.

stops 13–25% on the same data, and published classifier defenses report 67% at best. The defect also has a straightforward fix on the flatte

Why it matters: Baselines are far weaker than the code-level check.

C6 -- 2610.29869

Laundering still succeeds 100% of the time at every context size.

The laundering still succeeds 100% of the time at every size, and the executed rate, near-100% on the short transcripts, settles to ~20–40% in a full window. Th

Why it matters: Size does not attenuate the attack.

C7 -- 2610.29869

Executed rate settles to ~20–40% in a full window.

settles to ~20–40% in a full window. The short runs therefore measure the attack at its most favorable, an upper bound on the exec

Why it matters: Real agents compact at this size, so this is the realistic case.

T4: What raw evidence must be preserved

R3 -- 2610.30755

Mem0's UPDATE+DELETE account for only 4.39% of 33,902 LoCoMo write decisions.

together account for only 4.39% of the 33,902 write decisions on LoCoMo and 2.46% of the 428,858 write decisions on LongMemEval (Figure 2);

Why it matters: Mutation is rare in practice -- most proposed writes are duplicates or malformed.

R4 -- 2610.30755

Replacing UPDATE/DELETE with append-only ADD/NONE raises LoCoMo accuracy 67.36% -> 69.42%.

Replacing UPDATE with append-only ADD and DELETE with NONE raises LoCoMo accuracy from 67.36% to 69.42%, without reducing any category score (Table 2). Explicit mutation

Why it matters: Explicit mutation gives no accuracy gain; the simpler store is better.

R5 -- 2610.30755

Removing reranking lowers MemoryAgentBench Total by 3.92 points.

Removing reranking lowers Total by 3.92 points (pHolm=0.0012p_{\mathrm{Holm}}=0.0012), while returning only one item lowers it by 11.73 points. Final-kk values of 5, 10, 30, and 60 stay with

Why it matters: Reranking is the one component worth keeping.

R6 -- 2610.30755

Returning only one item lowers Total by 11.73 points.

while returning only one item lowers it by 11.73 points. Final-kk values of 5, 10, 30, and 60 stay within 1.3 points of the baseli

Why it matters: Retrieval depth matters far more than chunk size.

R7 -- 2610.30755

Top-120 retrieval loses 14.57 points.

Top-120 loses 14.57 points, driven by long-range understanding and conflict resolution. Chunk sizes from 128 to 1,024 do not signi

Why it matters: Driven by long-range understanding and conflict resolution.

M6 -- 2610.29823

On the 88 questions without photo evidence, Pi scores 100%.

On the 88 questions without photo evidence, Pi scores 100%, and full context, Mem0, Codex and Claude Code 99% each (Table 4). With the same input, then, Pi matc

Why it matters: Pi matches full context once captions are removed from the comparison.

T5: Interaction preferences are a different problem

P1 -- 2610.30485

MemPABench has four simulated users holding 14 preferences.

Four grounded simulated users hold 14 preferences across work and personal contexts. Their feedback spans 99 sessions over four months, with

Why it matters: Preferences span work and personal contexts.

P2 -- 2610.30485

Feedback spans 99 sessions over four months.

Their feedback spans 99 sessions over four months, with five event-driven changes per user testing memory updates. We evaluate pre

Why it matters: Five event-driven preference changes per user.

P3 -- 2610.30485

No Memory reaches 49.6% PSA against 45.2% random.

No Memory, where the PA selects by its defaults, reaches 49.6%, compared with 45.2% for random selection (Appendix A.1). Only File memory adds a clear gain. It raises PSA to 59.2%, a mean ga

Why it matters: Defaults alone nearly match chance-plus-margin.

P4 -- 2610.30485

File memory raises PSA to 59.2%, +9.6 points over No Memory.

It raises PSA to 59.2%, a mean gain of 9.6 points over No Memory (95% cluster-bootstrap interval [+3.9, +15.6]). PERMA (Liu et al., 2026), a benchmark of content preferen

Why it matters: The only condition with a clear gain; CI [+3.9, +15.6].

P5 -- 2610.30485

Mem0, MemOS and Simple RAG stay at No Memory level (PSA 50.0/50.0/49.9).

with PSA of 50.0%, 50.0%, and 49.9%. The intervals of all conditions and their comparisons with No Memory on each user–backbone pa

Why it matters: Retrieval-based memory does not beat having none.

P6 -- 2610.30485

File memory PET is 25.4%, more than twice No Memory's 11.2%.

Its PET is 25.4%, more than twice the 11.2% of No Memory, and it is higher on every backbone. However, the CS does not improve, at 14.4% against 21.8% (Appendix G.3). In

Why it matters: But its CS does not improve (14.4% vs 21.8%): context is not tracked.

P7 -- 2610.30485

Oracle lifts correct selection 33%->96% but BRS>=4 only 24%->59%.

Oracle raises correct selections of the five settings from 33% to 96%, but the share of replies with a BRS of 4 or higher rises only from 24% to 59%. Even with the IX Wrapper, the PA still writes too much. Afte

Why it matters: Knowing the preference does not override the model's interaction habits.

P8 -- 2610.30485

Information Flow is the hardest dimension, PSA 34–44%.

Information Flow is the hardest under every condition (Figure 8). Its PSA stays at 34–44%, the lowest of the four, while Expression Style (57–69%) and Disclosur

Why it matters: Expression Style (57-69%) and Disclosure (55-58%) are easier.

P9 -- 2610.30485

Anonymizing users drops shortcut accuracy 53.2% -> 46.4%, near the 45.2% chance.

reaches 53.2% accuracy from the source identity but 46.4% from the anonymized profile, close to the 45.2% chance level (Figure 10; Appendix

Why it matters: Confirms the benchmark measures memory, not name-guessing.

Cross-paper synthesis

Four of the five papers converge on one answer: the elaborate part of agent memory is not where the value is.

• Raw-RAG wins most settings (2610.30755) -- naive dense retrieval over raw history ranks first in six of seven benchmark-model settings, and no added memory-management operation improves on it. The one setting where it loses is the one where full context is affordable.

• Mutation is nearly free to drop (2610.30755) -- UPDATE+DELETE are 4.39% of LoCoMo writes and 2.46% of LongMemEval writes; replacing them with append-only ADD/NONE raises accuracy 67.36% -> 69.42%.

• Agent memory is a patch for a weak model (2610.29823) -- Mem0 helps GPT-4o-mini (78% vs 74%) but not GPT-6 Luna or Claude Opus 5.5, and at the top end Pi scores 99% for $6.95 against full context's 95% for $17.84.

• A stronger reader does not recover what a weaker writer left out (2610.29823) -- on Zep's GPT-4o-mini graph the three readers score 43/43/45%; on the graph Claude Opus 5.5 wrote they score 66/71/87%. Memory has a ceiling set by its writer.

• What must be preserved is raw evidence, not extracted facts (2610.30755, 2610.29823) -- verbatim chunks beat lossy artifacts; on the 88 questions with no photo evidence, Pi's lead over full context vanishes (100% vs 99%).

• Compaction destroys provenance deterministically (2610.29869) -- a forged approval launders into memory 87-100% of the time in flattening harnesses and 0% in role-typed ones; no prompt-level defense restores it, only a code-level check (77% catch, 2% collateral).

• Interaction preferences are a harder problem than factual recall (2610.30485) -- retrieval-based memory does not beat No Memory (PSA 50.0/50.0/49.9 vs 49.6%), and even an Oracle that hands the agent the right preference lifts delivered behavior only 24% -> 59%.

• The counter-example is training the memory manager itself (2610.29671) -- Uni-Mem jointly learns construction and retrieval and reaches 88.80% JJ, beating Memory-R2 (OSS) by 10.93 points and off-the-shelf Qwen3.5-27B (87.05%). The gain comes from the training scheme (Stage-Decoupled GRPO + ECA), not from a bigger model.

Where the five disagree

2610.30755 and 2610.29823 both conclude that raw-history search is competitive, but they disagree on why the alternative fails: 2610.30755 blames the rarity of useful mutation (4.39% of writes), while 2610.29823 blames the writer's ceiling. These are compatible -- a weak writer produces few useful mutations and a lossy store -- but neither paper measures the other's mechanism, so the dossier records the disagreement rather than resolving it.

2610.30485 is the outlier on the headline question: it studies interaction preferences, not factual recall, and finds memory nearly useless there. That is a different task family, not a refutation of the other four.

Claim ledger

44 claims across 5 papers (2610.29671:9, 2610.29823:9, 2610.29869:7, 2610.30485:9, 2610.30755:10).

id	paper	theme	value
R1	2610.30755	T1	six of seven
R2	2610.30755	T1	0.13–5.53 points
R3	2610.30755	T4	4.39%
R4	2610.30755	T4	67.36% to 69.42%
R5	2610.30755	T4	3.92 points
R6	2610.30755	T4	11.73 points
R7	2610.30755	T4	14.57 points
R8	2610.30755	T1	48.32%
R9	2610.30755	T1	47.68% to 15.37%
R10	2610.30755	T2	19.98 points
P1	2610.30485	T5	14 preferences
P2	2610.30485	T5	99 sessions over four months
P3	2610.30485	T5	49.6%
P4	2610.30485	T5	59.2%
P5	2610.30485	T5	50.0%, 50.0%, and 49.9%
P6	2610.30485	T5	25.4%
P7	2610.30485	T5	33% to 96%
P8	2610.30485	T5	34–44%
P9	2610.30485	T5	53.2%
M1	2610.29823	T2	419 turns, 152 questions
M2	2610.29823	T2	78%
M3	2610.29823	T2	92%
M4	2610.29823	T2	Pi 99%, Codex 98% and Claude Code 97%
M5	2610.29823	T2	$6.95
M6	2610.29823	T4	88 questions without photo evidence
M7	2610.29823	T2	10,000 input tokens
M8	2610.29823	T3	43%, 43% and 45%
M9	2610.29823	T3	151 questions
C1	2610.29869	T3	87–100%
C2	2610.29869	T3	0% of the time
C3	2610.29869	T3	~50%
C4	2610.29869	T3	77%
C5	2610.29869	T3	13–25%
C6	2610.29869	T3	100%
C7	2610.29869	T3	~20–40%
U1	2610.29671	T1	88.80%
U2	2610.29671	T1	10.93 percentage points
U3	2610.29671	T1	80.99%
U4	2610.29671	T1	1.29 and 3.03 percentage points
U5	2610.29671	T1	2.48 percentage points
U6	2610.29671	T1	8.18 percentage points
U7	2610.29671	T1	83.75%
U8	2610.29671	T2	76.58% to 88.80%
U9	2610.29671	T2	87.05%